- Email: [email protected]
Risk on using logistic regression to illustrate exposure-response relationship of infectious disease
A202
VA L U E I N H E A LT H 1 7 ( 2 0 1 4 ) A 1 - A 2 9 5
PRM121 Efficient estimation of the incremental cost-effectiveness ratio (ICER) using a new perspective Skrepnek G H 1, Sahai A 2 1The University of Oklahoma Health Sciences Center, Oklahoma City, OK, USA, 2The University of The West Indies, St. Augustine, Trinidad and Tobago .
.
.
Objectives: To develop and test a new method of estimating the ICER utilizing the harmonic mean. Methods: A statistically-efficient point and 95% confidence interval (CI) estimator of the ICER was derived that utilized the harmonic mean of costs and effects, hence applying the inverse of the mean of the inverses for use in statistical summarization. An additional correction factor developed through computational intelligence was also incorporated to capture existing information from the usual bootstrap ICER estimator. A simulation study of 1111 replications with 999 bootstrap resamples each utilizing Matlab R2012b was undertaken across illustrative positive and negative correlation structures of costs and outcomes for varying sample sizes of treatment and referent groups. Results were presented as relative efficiencies for point estimators, while coverage probability, coverage error, length, left and right bias, and relative bias were presented for the 95% CI. Results: Compared to the usual bootstrap approach, optimal methods based upon the harmonic mean yielded point estimates with greater relative efficiency across all analytic scenarios, ranging from 103.22% to 111.03%. The 95% CI coverage error was also consistently lower, deviating from the population value by 0.0005 to 0.0257 versus the usual bootstrap range of 0.0031 to 0.0302. Thus, an improved shortening of the CI length was found across all cases. The maximum relative bias of the new estimator was 0.7714 versus 0.2703, which reflected a somewhat higher left bias and lower right bias among positive correlation structures, and a greater right bias and lesser left bias among negative correlation structures. Conclusions: The new approach to estimate the ICER that utilized the harmonic mean allowed for more statistically-efficient point estimation. The 95% CIs presented with less coverage error and shorter lengths, though typically at the cost of a potential increase in relative bias. PRM122 Survival crossover adjustment and cost effectiveness analysis: An empirical and methodological review with application Hopkins R B 1, Campbell K 1, Burke N 1, Levine M 1, Thabane L 1, Duong M 2, Shum D 2, Goeree R 1 1McMaster University, Hamilton, ON, Canada, 2Hoffmann-La Roche Limited, Mississauga, ON, Canada .
.
.
.
.
.
.
.
.
Objectives: First, to summarize the methodological literature on the correction of overall survival for the impact of crossover. Second, to examine and compare the use of these statistical methods in cost-effectiveness analyses (CEAs). Third, to apply recommended statistical methods to correct overall survival in a clinical trial that included crossover. Methods: Medline, Embase, NHSEED and HEED databases, along with grey literature were searched for methodological evidence on the appropriate use of survival adjustment, including but not limited to, Inverse Probability of Censor Weighting (IPCW) and Rank Preserving Structural Failure Time Modelling (RPSFT). In addition, empirical CEAs that applied survival adjustment were identified and reviewed. The appropriate methods were applied to a trial with crossover to compare IPCW, RPSFT, intention to treat (ITT) and per protocol (PP) analysis. Results: The choice of IPCW or RPSFT depends on six factors: common treatment effect, true treatment effect, crossover percentage, disease severity, time dependence of treatment effect, and crossover mechanism. Nine placebo-controlled CEAs that applied survival adjustment were identified: two studies used one method without comparison, one study incorporated censor weighting for a meta-analysis, five studies reported one method and compared to either ITT or PP analysis, one as the primary analysis and four as sensitivity analysis. Only one study reported a comparison of multiple methods, IPCW and RPSFT. Empirically, PP, RPSFT and IPCW produce lower hazard ratios than ITT. Ranking of PP, RPSFT and IPCW varied by the factors. None of the six factors were discussed thoroughly in the empirical results. Based on the trial, all patients that crossed-over survived which violates the assumptions of common treatment effect for RPSFT and different disease severity for IPCW and RPSFT. Conclusions: Applying the six factors guides the a priori assessment of appropriate choice of crossover method. In this case, neither IPCW or RPSFT were appropriate. PRM123 A microsoft-excel based tool for running and critically appraising simple network meta-analyses using winbugs – an overview and application Brown S T 1, Cameron C 2, Grima D T 1, Wells G 3 1Cornerstone Research Group, Burlington, ON, Canada, 2University of Ottawa, Ottawa, ON, Canada, 3University of Ottawa Heart Institute, Ottawa, ON, Canada .
.
.
.
.
.
Objectives: The role of network meta-analyses has increased dramatically in recent years. WinBUGS has been the most widely used software to conduct network meta-analyses. However, the learning curve for using WinBUGS to conduct network meta-analyses successfully can be daunting, especially for new users. Further, critically appraising network meta-analyses conducted in WinBUGS is challenging given the limited data analysis and graphical output fromWinBUGS, thus network meta-analyses often rely on different software packages. The objective is to develop a tool which 1.) makes running network meta-analyses more accessible to novice WinBUGS users; and 2.) facilitates a more transparent and efficient critical appraisal of network meta-analyses. Methods: We developed a freely available Microsoft-Excel based tool, programmed in Visual Basic for applications within Excel, which provides an interface for conducting a network meta-analysis using WinBUGS from within Microsoft Excel. This tool allows the user to modify assumptions and to run the network meta-analysis, and results are returned to an Excel spreadsheet. The tool displays the data, evidence networks, forest plots, rankograms, and inconsistency plots all entirely within Microsoft Excel. Results: We demon-
strate the application of our freely available Microsoft-Excel based tool using an example of a network meta-analysis of anti-platelet agents in patients scheduled for percutaneous coronary inteventions. Conclusions: Use of this freely available Microsoft-Excel based tool successfully demonstrated its ability to make running network meta-analyses more accessible to novice WinBUGS users, and facilitate more transparent critical appraisal of network meta-analyses. PRM124 Matching with multiple control groups to maximize use of registry data from patients with schizophrenia Lopatto J 1, Song X 2, Juneau P 3, Benson C 1, Olson W H 4, Fastenau J 1 Scientific Affairs, LLC, Titusville, NJ, USA, 2Truven Health Analytics, Cambridge, MA, USA, 3Truven Health Analytics, Boyds, MD, USA, 4Janssen Pharmaceuticals, Inc., Titusville, NJ, USA .
.
.
.
.
.
.
1Janssen
Objectives: In order to examine comparative research questions using data from a naturalistic, observational study (REACH-OUT) of adult patients with schizophrenia, non-traditional methods were needed. This abstract describes the use of multiple control groups to maximize inclusion of paliperidone palmitate (PP) patients in REACH-OUT who would otherwise be excluded from analysis due to poor propensity score matching with registry controls (patients receiving oral atypical antipsychotics (OAT)). The matched cohorts (PP and combined OAT) will be used in future resource use comparisons. Methods: Because matching on propensity of PP treatment did not yield an adequate number of matched registry controls, a secondary set of OAT controls was extracted from MarketScan® claims data. PP patients unmatched to registry controls were matched to claims controls using a 1:1 propensity score matching. Post-match baseline characteristics for the PP and combined OAT cohort were examined using descriptive statistics. Outcomes from the claims-based control group will be adjusted for source bias based on 500 simulations, according to the Stuart-Rubin (S-R) methodology. Results: Out of 354 PP with non-missing baseline data, 190 were matched to registry controls and the remaining 164 were matched to the supplemental control group. The final matched PP and OAT cohorts were balanced in observed attributes such as age (41.4 years vs. 42.0, p= 0.552), gender distribution (70.3% vs. 65.5% male, p= 0.171), ≥ 1 baseline hospitalization (29.7% vs. 34.5%, p= 0.171), and ≥ 1 baseline ER visit (31.1% vs. 35.6%, P= 0.202), respectively. Initial S-R simulations suggest outcomes for the supplemented control group are similar to the cohort of all REACH-OUT controls (e.g., 6 month admission rates, 18.9% and 22.6%, respectively). Conclusions: Use of multiple control groups permitted successful propensity score matching of all registry PP patients allowing for greater power, better precision, and increased external validity in the forthcoming analysis of the treatment effect of PP.
Research on Methods – Study Design PRM125 Risk on using logistic regression to illustrate exposure-response relationship of infectious disease Ren J , Asche C , Kirkness C S University of Illinois, Peoria, IL, USA .
.
.
.
Objectives: Logistic regression is widely used to assess the likelihood of an infectious disease as a function of a risk or exposure factor (and covariates), to illustrate the exposure-response relationship. However, because the exposure of patients with infectious disease is an intricate net instead of independent factors, the statistical power of logistic regression may be compromised leading to an inaccurate conclusion. Therefore, this study aims to examine the statistical power of logistic regression using simulated data of infectious disease. Methods: A dynamic human immunodeficiency virus (HIV) infection was simulated among 10000 individuals with 1% initial prevalence and 7% target prevalence. Monte Carlo simulation method was used to examine the statistical powers of regular logistic regression (grouping sexual partners into 0-2, 3-5, 6-8, 9-11, 12-14 and ≥ 15), transformed logistic regression (using log[1+number of sexual partners]) and negative binomial regression on estimating the risk of HIV infection along with increasing number of unprotected sexual partners. Results: Regular logistic regression had poor statistical power and overestimated the odds ratio (OR) when the number of sex partner was more than 11 (power was 78% for 12-14 partners and only 4% for over 14 partners). Transformed logistic regression overstated the odds ratio even while the number of sex partner was small. Negative binomial regression had 100% power finding the association between HIV infection and the number of sexual partners, yet it was not available to provide odds ratio. Conclusions: Due to the diverse distribution of exposure in infectious disease (negative binomial), evaluations that include logistic regression, to explore the exposure-response relationship, provide improved rigor to base decisions. PRM126 Burden of narcolepsy disease (bond) study: Validation of using a single diagnosis code to define presence of an orphan condition in medical claims data Villa K 1, Reaven N 2, Funk S 2, McGaughey K 3, Ohayon M 4, Guilleminault C 4, Ruoff C 4, Black J 4 1Jazz Pharmaceuticals, Palo Alto, CA, USA, 2Strategic Health Resources, La Canada, CA, USA, 3Cal Poly State University, San Luis Obispo, CA, USA, 4Stanford University Center for Sleep Research and Medicine, Redwood City, CA, USA .
.
.
.
.
.
.
.
Objectives: A US medical claims-based analysis was designed to evaluate burden of illness associated with narcolepsy, a chronic, non-progressive disease often presenting early in life. Because diagnostic testing for narcolepsy is not normally repeated, objective evidence of a diagnosis would be absent for many patients in a time-limited data set. Therefore, internal validation of a study population selected using diagnosis codes was performed. Methods: Within the 5-year data collection period (2006 through 2010), 9312 continuously insured adult patients were
VA L U E I N H E A LT H 1 7 ( 2 0 1 4 ) A 1 - A 2 9 5
PRM121 Efficient estimation of the incremental cost-effectiveness ratio (ICER) using a new perspective Skrepnek G H 1, Sahai A 2 1The University of Oklahoma Health Sciences Center, Oklahoma City, OK, USA, 2The University of The West Indies, St. Augustine, Trinidad and Tobago .
.
.
Objectives: To develop and test a new method of estimating the ICER utilizing the harmonic mean. Methods: A statistically-efficient point and 95% confidence interval (CI) estimator of the ICER was derived that utilized the harmonic mean of costs and effects, hence applying the inverse of the mean of the inverses for use in statistical summarization. An additional correction factor developed through computational intelligence was also incorporated to capture existing information from the usual bootstrap ICER estimator. A simulation study of 1111 replications with 999 bootstrap resamples each utilizing Matlab R2012b was undertaken across illustrative positive and negative correlation structures of costs and outcomes for varying sample sizes of treatment and referent groups. Results were presented as relative efficiencies for point estimators, while coverage probability, coverage error, length, left and right bias, and relative bias were presented for the 95% CI. Results: Compared to the usual bootstrap approach, optimal methods based upon the harmonic mean yielded point estimates with greater relative efficiency across all analytic scenarios, ranging from 103.22% to 111.03%. The 95% CI coverage error was also consistently lower, deviating from the population value by 0.0005 to 0.0257 versus the usual bootstrap range of 0.0031 to 0.0302. Thus, an improved shortening of the CI length was found across all cases. The maximum relative bias of the new estimator was 0.7714 versus 0.2703, which reflected a somewhat higher left bias and lower right bias among positive correlation structures, and a greater right bias and lesser left bias among negative correlation structures. Conclusions: The new approach to estimate the ICER that utilized the harmonic mean allowed for more statistically-efficient point estimation. The 95% CIs presented with less coverage error and shorter lengths, though typically at the cost of a potential increase in relative bias. PRM122 Survival crossover adjustment and cost effectiveness analysis: An empirical and methodological review with application Hopkins R B 1, Campbell K 1, Burke N 1, Levine M 1, Thabane L 1, Duong M 2, Shum D 2, Goeree R 1 1McMaster University, Hamilton, ON, Canada, 2Hoffmann-La Roche Limited, Mississauga, ON, Canada .
.
.
.
.
.
.
.
.
Objectives: First, to summarize the methodological literature on the correction of overall survival for the impact of crossover. Second, to examine and compare the use of these statistical methods in cost-effectiveness analyses (CEAs). Third, to apply recommended statistical methods to correct overall survival in a clinical trial that included crossover. Methods: Medline, Embase, NHSEED and HEED databases, along with grey literature were searched for methodological evidence on the appropriate use of survival adjustment, including but not limited to, Inverse Probability of Censor Weighting (IPCW) and Rank Preserving Structural Failure Time Modelling (RPSFT). In addition, empirical CEAs that applied survival adjustment were identified and reviewed. The appropriate methods were applied to a trial with crossover to compare IPCW, RPSFT, intention to treat (ITT) and per protocol (PP) analysis. Results: The choice of IPCW or RPSFT depends on six factors: common treatment effect, true treatment effect, crossover percentage, disease severity, time dependence of treatment effect, and crossover mechanism. Nine placebo-controlled CEAs that applied survival adjustment were identified: two studies used one method without comparison, one study incorporated censor weighting for a meta-analysis, five studies reported one method and compared to either ITT or PP analysis, one as the primary analysis and four as sensitivity analysis. Only one study reported a comparison of multiple methods, IPCW and RPSFT. Empirically, PP, RPSFT and IPCW produce lower hazard ratios than ITT. Ranking of PP, RPSFT and IPCW varied by the factors. None of the six factors were discussed thoroughly in the empirical results. Based on the trial, all patients that crossed-over survived which violates the assumptions of common treatment effect for RPSFT and different disease severity for IPCW and RPSFT. Conclusions: Applying the six factors guides the a priori assessment of appropriate choice of crossover method. In this case, neither IPCW or RPSFT were appropriate. PRM123 A microsoft-excel based tool for running and critically appraising simple network meta-analyses using winbugs – an overview and application Brown S T 1, Cameron C 2, Grima D T 1, Wells G 3 1Cornerstone Research Group, Burlington, ON, Canada, 2University of Ottawa, Ottawa, ON, Canada, 3University of Ottawa Heart Institute, Ottawa, ON, Canada .
.
.
.
.
.
Objectives: The role of network meta-analyses has increased dramatically in recent years. WinBUGS has been the most widely used software to conduct network meta-analyses. However, the learning curve for using WinBUGS to conduct network meta-analyses successfully can be daunting, especially for new users. Further, critically appraising network meta-analyses conducted in WinBUGS is challenging given the limited data analysis and graphical output fromWinBUGS, thus network meta-analyses often rely on different software packages. The objective is to develop a tool which 1.) makes running network meta-analyses more accessible to novice WinBUGS users; and 2.) facilitates a more transparent and efficient critical appraisal of network meta-analyses. Methods: We developed a freely available Microsoft-Excel based tool, programmed in Visual Basic for applications within Excel, which provides an interface for conducting a network meta-analysis using WinBUGS from within Microsoft Excel. This tool allows the user to modify assumptions and to run the network meta-analysis, and results are returned to an Excel spreadsheet. The tool displays the data, evidence networks, forest plots, rankograms, and inconsistency plots all entirely within Microsoft Excel. Results: We demon-
strate the application of our freely available Microsoft-Excel based tool using an example of a network meta-analysis of anti-platelet agents in patients scheduled for percutaneous coronary inteventions. Conclusions: Use of this freely available Microsoft-Excel based tool successfully demonstrated its ability to make running network meta-analyses more accessible to novice WinBUGS users, and facilitate more transparent critical appraisal of network meta-analyses. PRM124 Matching with multiple control groups to maximize use of registry data from patients with schizophrenia Lopatto J 1, Song X 2, Juneau P 3, Benson C 1, Olson W H 4, Fastenau J 1 Scientific Affairs, LLC, Titusville, NJ, USA, 2Truven Health Analytics, Cambridge, MA, USA, 3Truven Health Analytics, Boyds, MD, USA, 4Janssen Pharmaceuticals, Inc., Titusville, NJ, USA .
.
.
.
.
.
.
1Janssen
Objectives: In order to examine comparative research questions using data from a naturalistic, observational study (REACH-OUT) of adult patients with schizophrenia, non-traditional methods were needed. This abstract describes the use of multiple control groups to maximize inclusion of paliperidone palmitate (PP) patients in REACH-OUT who would otherwise be excluded from analysis due to poor propensity score matching with registry controls (patients receiving oral atypical antipsychotics (OAT)). The matched cohorts (PP and combined OAT) will be used in future resource use comparisons. Methods: Because matching on propensity of PP treatment did not yield an adequate number of matched registry controls, a secondary set of OAT controls was extracted from MarketScan® claims data. PP patients unmatched to registry controls were matched to claims controls using a 1:1 propensity score matching. Post-match baseline characteristics for the PP and combined OAT cohort were examined using descriptive statistics. Outcomes from the claims-based control group will be adjusted for source bias based on 500 simulations, according to the Stuart-Rubin (S-R) methodology. Results: Out of 354 PP with non-missing baseline data, 190 were matched to registry controls and the remaining 164 were matched to the supplemental control group. The final matched PP and OAT cohorts were balanced in observed attributes such as age (41.4 years vs. 42.0, p= 0.552), gender distribution (70.3% vs. 65.5% male, p= 0.171), ≥ 1 baseline hospitalization (29.7% vs. 34.5%, p= 0.171), and ≥ 1 baseline ER visit (31.1% vs. 35.6%, P= 0.202), respectively. Initial S-R simulations suggest outcomes for the supplemented control group are similar to the cohort of all REACH-OUT controls (e.g., 6 month admission rates, 18.9% and 22.6%, respectively). Conclusions: Use of multiple control groups permitted successful propensity score matching of all registry PP patients allowing for greater power, better precision, and increased external validity in the forthcoming analysis of the treatment effect of PP.
Research on Methods – Study Design PRM125 Risk on using logistic regression to illustrate exposure-response relationship of infectious disease Ren J , Asche C , Kirkness C S University of Illinois, Peoria, IL, USA .
.
.
.
Objectives: Logistic regression is widely used to assess the likelihood of an infectious disease as a function of a risk or exposure factor (and covariates), to illustrate the exposure-response relationship. However, because the exposure of patients with infectious disease is an intricate net instead of independent factors, the statistical power of logistic regression may be compromised leading to an inaccurate conclusion. Therefore, this study aims to examine the statistical power of logistic regression using simulated data of infectious disease. Methods: A dynamic human immunodeficiency virus (HIV) infection was simulated among 10000 individuals with 1% initial prevalence and 7% target prevalence. Monte Carlo simulation method was used to examine the statistical powers of regular logistic regression (grouping sexual partners into 0-2, 3-5, 6-8, 9-11, 12-14 and ≥ 15), transformed logistic regression (using log[1+number of sexual partners]) and negative binomial regression on estimating the risk of HIV infection along with increasing number of unprotected sexual partners. Results: Regular logistic regression had poor statistical power and overestimated the odds ratio (OR) when the number of sex partner was more than 11 (power was 78% for 12-14 partners and only 4% for over 14 partners). Transformed logistic regression overstated the odds ratio even while the number of sex partner was small. Negative binomial regression had 100% power finding the association between HIV infection and the number of sexual partners, yet it was not available to provide odds ratio. Conclusions: Due to the diverse distribution of exposure in infectious disease (negative binomial), evaluations that include logistic regression, to explore the exposure-response relationship, provide improved rigor to base decisions. PRM126 Burden of narcolepsy disease (bond) study: Validation of using a single diagnosis code to define presence of an orphan condition in medical claims data Villa K 1, Reaven N 2, Funk S 2, McGaughey K 3, Ohayon M 4, Guilleminault C 4, Ruoff C 4, Black J 4 1Jazz Pharmaceuticals, Palo Alto, CA, USA, 2Strategic Health Resources, La Canada, CA, USA, 3Cal Poly State University, San Luis Obispo, CA, USA, 4Stanford University Center for Sleep Research and Medicine, Redwood City, CA, USA .
.
.
.
.
.
.
.
Objectives: A US medical claims-based analysis was designed to evaluate burden of illness associated with narcolepsy, a chronic, non-progressive disease often presenting early in life. Because diagnostic testing for narcolepsy is not normally repeated, objective evidence of a diagnosis would be absent for many patients in a time-limited data set. Therefore, internal validation of a study population selected using diagnosis codes was performed. Methods: Within the 5-year data collection period (2006 through 2010), 9312 continuously insured adult patients were